A method and system for suppressing strong interference in controlled-source electromagnetic data

Abstract

This application belongs to the cross - field of electronic information, artificial intelligence and geophysics. Specifically, it discloses a method and system for suppressing strong interference in controlled - source electromagnetic data. The method is as follows: Input the to - be - processed controlled - source electromagnetic data into the controlled - source electromagnetic data classification model to obtain the first noise - free controlled - source electromagnetic data and the noise - containing controlled - source electromagnetic data; Convert the noise - containing controlled - source electromagnetic data into two - dimensional image data through a dimension conversion function and input it into the DWTSC - UNet denoising network model to obtain the denoising data result; Use the dimension conversion function to perform an inverse dimension conversion on the denoising data result to obtain the second noise - free controlled - source electromagnetic data; Combine the first noise - free controlled - source electromagnetic data and the second noise - free controlled - source electromagnetic data to obtain the complete noise - free controlled - source electromagnetic data. This application can suppress strong interference with high precision and high efficiency.

Description

Technical Field

[0001] This application belongs to the cross - field of electronic information, artificial intelligence and geophysics. More specifically, it relates to a method and system for suppressing strong interference of controlled - source electromagnetic data (Controlled - Source Electromagnetic Method, CSEM) based on a SC - UNet (Discrete Wavelet Transform optimized Spatial and Channel Attention Mechanism U - shaped Neural Network, DWTSC - UNet) network optimized by dimension conversion and discrete wavelet transform. Background Technique

[0002] The controlled - source electromagnetic method is an electromagnetic sounding method widely used in fields such as near - surface geophysical exploration and engineering survey. Currently, it has been increasingly widely used in fields such as shale gas exploration, metal ore exploration, and engineering geophysical exploration. Its essence is to deduce the distribution of underground media by using the different responses generated by the electrical property structure differences of underground media. The controlled - source electromagnetic method has a known frequency and a controllable field source, and has a higher anti - interference ability compared with the natural - field - source electromagnetic method. With the continuous increase of the proportion of urbanization and industrialization, the distribution of human - made noise is becoming wider. Since the controlled - source electromagnetic signal is extremely vulnerable to human - made noise, with the continuous improvement of the requirements for exploration depth and accuracy, suppressing the noise in the controlled - source electromagnetic signal has become increasingly important.

[0003] Facing the severe human - made noise pollution, the common practice is the data screening method. Given the periodic characteristics of the controlled - source electromagnetic signal, the simplest method is to screen and eliminate the segments containing strong human - made interference in units of periods. The data screening method can improve the data quality to a certain extent and extract effective signals. However, there are still its limitations: First, the traditional manual selection of high - quality data not only requires certain prior knowledge, but also has low efficiency and is prone to subjective factors; Second, a single data statistical feature is difficult to distinguish noise segments and effective signals in a complex noise environment; if multiple groups of data statistical features are used, the mutual influence between feature parameters needs to be considered; Third, there is still residual noise (such as harmonics and Gaussian white noise, etc.) in the screened data.

[0004] In recent years, deep learning technology has developed rapidly and is increasingly widely used in the electromagnetic field. The application of deep learning algorithms in electromagnetic signal processing has gradually increased. However, due to the complexity and variability of controlled-source electromagnetic signals, it is difficult for the learning model to accurately capture the complete features of useful signals in the case of strong noise interference, and overfitting or underfitting may occur. In addition, the deep learning networks adopted by existing methods have been proposed for decades, and there are problems such as vanishing or exploding gradients, excessive number of parameters, excessive memory occupation, and the need to improve denoising accuracy. Summary of the Invention

[0005] Aiming at the defects of the prior art, the purpose of this application is to provide a method and system for suppressing strong interference of controlled-source electromagnetic data, aiming to solve the problem that due to the complexity and polygon of controlled-source electromagnetic signals, it is difficult for the learning model to accurately capture the complete features of useful signals in the case of strong noise interference, and overfitting or underfitting may occur, which may lead to insufficient denoising accuracy.

[0006] To achieve the above purpose, in the first aspect, this application provides a method for suppressing strong interference of controlled-source electromagnetic data, including the following steps:

[0007] Input the to-be-processed controlled-source electromagnetic data into the controlled-source electromagnetic data classification model to obtain the first noise-free controlled-source electromagnetic data and the noise-containing controlled-source electromagnetic data;

[0008] Convert the noise-containing controlled-source electromagnetic data into two-dimensional image data through the dimension conversion function and input it into the DWTSC-UNet denoising network model to obtain the denoising data result;

[0009] Perform inverse dimension conversion on the denoising data result using the dimension conversion function to obtain the second noise-free controlled-source electromagnetic data;

[0010] Combine the first noise-free controlled-source electromagnetic data and the second noise-free controlled-source electromagnetic data to obtain the complete noise-free controlled-source electromagnetic data;

[0011] Among them, the DWTSC-UNet denoising network model is a U-shaped structure network that combines discrete wavelet transform, spatial attention mechanism, and channel attention mechanism;

[0012] The discrete wavelet transform is used to provide high-frequency signals and low-frequency signals, perform weighted learning on different frequency components of the two-dimensional image data, and extract multi-scale features from the two-dimensional image data; the channel attention mechanism is used to weight different dimensions of the feature map; the spatial attention mechanism is used to weight different spatial positions of the feature map.

[0013] Further preferably, the training method of the DWTSC-UNet denoising network model includes the following steps:

[0014] Select a controlled-source electromagnetic data sample with noise from the measured data, and use the dictionary learning method for denoising to separate the measured noise signal;

[0015] Add the measured noise signal and the simulated noise signal to the noiseless controlled-source electromagnetic data to form the controlled-source electromagnetic data with noise, and form a sample pair with the noiseless controlled-source electromagnetic data sample to construct a one-dimensional denoising sample library;

[0016] Segment the one-dimensional denoising sample library, and use the dimension conversion function to perform dimension conversion on each sample to obtain a two-dimensional image sample library;

[0017] Input the two-dimensional image sample library into the DWTSC-UNet denoising network model for training.

[0018] Further preferably, the DWTSC-UNet denoising network model includes an input layer, a hidden layer, and an output layer; the hidden layer includes: a discrete wavelet transform layer, a first shape reconstruction layer, a first convolutional block, a first spatial attention mechanism module, a second convolutional block, a second spatial attention mechanism module, a third convolutional block, a first upsampling block, a second shape reconstruction layer, a first skip connection layer, a first channel attention mechanism layer, a second upsampling block, a second skip connection layer, a second channel attention mechanism module, and a third upsampling block.

[0019] Further preferably, the discrete wavelet transform layer is used to decompose the two-dimensional image data into a low-frequency approximation part, a horizontal high-frequency part, a vertical high-frequency part, and a diagonal high-frequency part; the low-frequency approximation part contains the basic contour of the two-dimensional image data; the horizontal high-frequency part contains the horizontal edge information of the two-dimensional image data; the vertical high-frequency part contains the vertical edge information of the two-dimensional image data; the diagonal high-frequency part contains the diagonal edge information of the two-dimensional image data.

[0020] Further preferably, the discrete wavelet transform layer uses the db5 wavelet basis type and the boundary processing mode of periodic boundary conditions.

[0021] Further preferably, the controlled-source electromagnetic data classification model is the IncepTCN classification network model.

[0022] In a second aspect, the present application provides a controlled-source electromagnetic data strong interference suppression system, including:

[0023] A data classification module, configured to input the to-be-processed controlled-source electromagnetic data into the controlled-source electromagnetic data classification model to obtain the first noiseless controlled-source electromagnetic data and the controlled-source electromagnetic data with noise;

[0024] A data denoising module, which is used to convert the noisy controlled-source electromagnetic data into two-dimensional image data through a dimension conversion function and input it into the DWTSC-UNet denoising network model to obtain the denoised data result;

[0025] A dimension conversion module, which is used to perform an inverse dimension conversion on the denoised data result by using a dimension conversion function to obtain the second noise-free controlled-source electromagnetic data;

[0026] A data combination module, which is used to combine the first noise-free controlled-source electromagnetic data and the second noise-free controlled-source electromagnetic data to obtain the complete noise-free controlled-source electromagnetic data;

[0027] Among them, the DWTSC-UNet denoising network model is a U-shaped structure network that combines discrete wavelet transform, spatial attention mechanism, and channel attention mechanism;

[0028] The discrete wavelet transform is used to provide high-frequency signals and low-frequency signals, perform weighted learning on different frequency components of the two-dimensional image data, and perform multi-scale feature extraction from the two-dimensional image data; the channel attention mechanism is used to weight different dimensions of the feature map; the spatial attention mechanism is used to weight different spatial positions of the feature map.

[0029] Further preferably, the controlled-source electromagnetic data strong interference suppression system further includes a denoising model training module, which includes:

[0030] An actual measured noise signal acquisition unit, which is used to select a noisy controlled-source electromagnetic data sample from the actual measured data, perform denoising by using the dictionary learning method, and separate the actual measured noise signal;

[0031] A one-dimensional denoising sample library construction unit, which is used to add the actual measured noise signal and the simulated noise signal to the noise-free controlled-source electromagnetic data to form the noisy controlled-source electromagnetic data, form a sample pair with the noise-free controlled-source electromagnetic data sample, and construct a one-dimensional denoising sample library;

[0032] A dimension conversion unit, which segments the one-dimensional denoising sample library and uses a dimension conversion function to perform dimension conversion on each sample to obtain a two-dimensional image sample library;

[0033] A model training unit, which is used to input the two-dimensional image sample library into the DWTSC-UNet denoising network model for training.

[0034] Further preferably, the DWTSC-UNet denoising network model in the data denoising module includes an input layer, a hidden layer, and an output layer; the hidden layer includes: a discrete wavelet transform layer, a first shape reconstruction layer, a first convolutional block, a first spatial attention mechanism module, a second convolutional block, a second spatial attention mechanism module, a third convolutional block, a first upsampling block, a second shape reconstruction layer, a first skip connection layer, a first channel attention mechanism layer, a second upsampling block, a second skip connection layer, a second channel attention mechanism module, and a third upsampling block;

[0035] Among them, the discrete wavelet transform layer is used to decompose the two-dimensional image data into a low-frequency approximation part, a horizontal high-frequency part, a vertical high-frequency part, and a diagonal high-frequency part; the low-frequency approximation part contains the basic contour of the two-dimensional image data; the horizontal high-frequency part contains the horizontal edge information of the two-dimensional image data; the vertical high-frequency part contains the vertical edge information of the two-dimensional image data; the diagonal high-frequency part contains the diagonal edge information of the two-dimensional image data.

[0036] Further preferably, the discrete wavelet transform layer uses a wavelet basis type of db5 wavelet basis and a boundary processing mode of periodic boundary conditions.

[0037] Further preferably, the controlled-source electromagnetic data classification model in the data classification module is an IncepTCN classification network model.

[0038] In a third aspect, the present application provides an electronic device, including: at least one memory for storing a program; at least one processor for executing the program stored in the memory, and when the program stored in the memory is executed, the processor is used to execute the method described in the first aspect or any possible implementation manner of the first aspect.

[0039] In a fourth aspect, the present application provides a computer-readable storage medium storing a computer program, and when the computer program runs on a processor, it causes the processor to execute the method described in the first aspect or any possible implementation manner of the first aspect.

[0040] In a fifth aspect, the present application provides a computer program product, and when the computer program product runs on a processor, it causes the processor to execute the method described in the first aspect or any possible implementation manner of the first aspect.

[0041] It can be understood that the beneficial effects of the above second aspect to fifth aspect can refer to the relevant descriptions in the first aspect above, and will not be elaborated here.

[0042] Generally speaking, compared with the prior art through the above technical solutions conceived by the present application, the following beneficial effects are achieved:

[0043] The strong interference suppression method for controlled-source electromagnetic data provided by this application introduces a dimension conversion method, converts the problem of denoising controlled-source electromagnetic signals into an image denoising problem, and uses the proposed signal-to-image transformation method to convert the one-dimensional controlled-source electromagnetic time series into two-dimensional data to retain the structural characteristics of the original controlled-source electromagnetic signals.

[0044] The strong interference suppression method for controlled-source electromagnetic data provided by this application introduces a method combining discrete wavelet transform and UNet network, making full use of the advantages of discrete wavelet transform and UNet architecture. First, the discrete wavelet transform performs multi-scale feature extraction on the controlled-source electromagnetic data, extracting different frequency information, which helps the network better understand the content of the controlled-source electromagnetic data. Then, combining the discrete wavelet transform and UNet architecture can perform more effective feature fusion between the encoder and decoder of the model, improving the performance of the network. In addition, UNet itself has good local feature capture ability. Using skip connections, the features extracted by the encoder part are directly passed to the decoder part, and the discrete wavelet transform can further strengthen this ability, especially at the detail level of the image, helping the model better capture the local structural information of the image. Moreover, the discrete wavelet transform can compress the high-dimensional representation of the image into a low-dimensional multi-scale representation, reducing the computational complexity and improving the processing efficiency of the network.

[0045] This application provides a strong interference suppression method for controlled-source electromagnetic data, which makes full use of the ability of the attention mechanism to improve the model's feature extraction ability, the multi-scale feature extraction ability of discrete wavelet transform, and the unique encoding-decoding structure, skip connections, and effective integration ability of local and global information of the UNet network, and proposes a new denoising network, namely the DWTSC-UNet denoising network. Without inspecting the effective signals, it can achieve strong noise suppression of controlled-source electromagnetic data, and the calculated apparent resistivity curve of the denoised data is significantly improved, becoming smoother and more continuous.

[0046] This application provides a strong interference suppression method for controlled-source electromagnetic data. After the model training is completed, the processes of data processing such as recognition and denoising are all automatically completed by the computer without any manual intervention, and there is no experience requirement for data processing operators. It not only eliminates the problem of subjective deviation caused by manually setting thresholds in traditional methods, but also improves the adaptability to different types of noise. Description of the Drawings

[0047] Figure 1 is one of the schematic flowcharts of the strong interference suppression method for controlled-source electromagnetic data provided by the embodiments of this application;

[0048] Figure 2 is the one-dimensional denoising sample library diagram provided by the embodiments of this application;

[0049] Figure 3 It is a schematic diagram of dimensionality conversion by the dimensionality conversion function provided in the embodiments of the present application;

[0050] Figure 4 It is an architecture diagram of the WDTSC-UNet denoising network model provided in the embodiments of the present application;

[0051] Figure 5 It is a schematic diagram of discrete wavelet transform data decomposition provided in the embodiments of the present application;

[0052] Figure 6 It is a graph showing the change of the loss curve during the training process of the WDTSC-UNet denoising network model provided in the embodiments of the present application;

[0053] Figure 7 It is a classification sample library diagram provided in the embodiments of the present application;

[0054] Figure 8 It is an architecture diagram of the IncepTCN classification network model provided in the embodiments of the present application;

[0055] Figure 9 It is a graph showing the change of accuracy and loss curve during the training process of the IncepTCN classification network model provided in the embodiments of the present application;

[0056] Figure 10 It is a confusion matrix diagram of the IncepTCN classification network model provided in the embodiments of the present application;

[0057] Figure 11 It is a classification effect diagram of the IncepTCN classification network model provided in the embodiments of the present application;

[0058] Figure 12 It is a denoising effect diagram of the DWTSC-UNet model provided in the embodiments of the present application;

[0059] Figure 13 It is a U / I curve and apparent resistivity curve diagram calculated from the noisy data and the denoised data provided in the embodiments of the present application. Detailed implementation manners

[0060] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0061] In this text, the term "and / or" describes the relationship between associated objects, indicating three possible relationships. For example, A and / or B can represent three cases: A exists alone, A and B exist simultaneously, and B exists alone. In this text, the symbol " / " indicates that the associated objects are in an "or" relationship. For example, A / B means A or B.

[0062] In the description of the specification and claims of this application, terms such as "first" and "second" are used to distinguish different objects, rather than to describe a specific order of the objects.

[0063] In the embodiments of this application, words such as "exemplary" or "for example" are used to give examples, illustrations, or explanations. Any embodiment or design solution described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Rather, the use of words such as "exemplary" or "for example" is intended to present relevant concepts in a specific manner.

[0064] In the description of the embodiments of this application, unless otherwise specified, the meaning of "a plurality of" refers to two or more.

[0065] The embodiments of this application will be described below with reference to the accompanying drawings in the embodiments of this application.

[0066] As Figure 1 shown, the embodiments of this application provide a method for suppressing strong interference in controlled-source electromagnetic data based on a SC-UNet network optimized by dimension conversion and discrete wavelet transform, specifically including the following steps:

[0067] Step 1: Select controlled-source electromagnetic data with high discrimination and containing typical noise from a large amount of measured data;

[0068] Step 2: Create a denoising sample library. Use the dictionary learning method to denoise the measured noisy samples, add the separated noise to high-quality controlled-source electromagnetic data to form partially noisy data samples. Another part of the noise samples is made by adding artificially synthesized simulated noise to high-quality data. Form sample pairs of the noisy controlled-source electromagnetic data samples and the corresponding high-quality controlled-source electromagnetic data samples to construct a one-dimensional denoising sample library;

[0069] Step 3: Segment the denoising sample library constructed in Step 2, and use the dimension conversion function to perform dimension conversion on each sample respectively, converting the one-dimensional data into a two-dimensional matrix, that is, obtaining a two-dimensional sample library;

[0070] Step 4: Construct a DWTSC-UNet denoising network model, input the two-dimensional sample library obtained in Step 3 into the DWTSC-UNet denoising network model for training to obtain a controlled-source electromagnetic data denoising model;

[0071] Step 5: Use each data segment in Step 1 as a sample and label the category of the sample with a class label to construct a classification sample library; the class label indicates whether the sample is a noisy data segment or a high-quality data segment.

[0072] Step 6: Construct an IncepTCN classification network, and input the classification sample library into the IncepTCN classification network for training to obtain a controlled-source electromagnetic data classification model.

[0073] Step 7: Input the to-be-processed controlled-source electromagnetic data, first segment it, and then input it into the controlled-source electromagnetic data classification model to obtain the classification result of each data segment, and retain the data segments identified as high-quality by the model.

[0074] Step 8: Input the controlled-source electromagnetic data segments identified as noisy by the classification model in Step 7 into the controlled-source electromagnetic data denoising model to obtain the data denoising result of each data segment.

[0075] Step 9: Use the dimension conversion function in Step 3 to perform inverse dimension conversion on the denoising result obtained in Step 8, and convert the two-dimensional image data into a one-dimensional time series.

[0076] Step 10: Combine the high-quality data retained in Step 7 and the data after dimension conversion of the denoised data in Step 9 to obtain a complete controlled-source electromagnetic denoising sequence.

[0077] Further preferably, the dimension conversion function is a data dimension transformation method. Since the length of the input one-dimensional data is 1600, it can be denoted as , and the target is to convert it into a 40×40 two-dimensional matrix A and fill the data in an "S" shape; the elements of the matrix can be expressed as ;

[0078] For odd rows i (such as the first row, the third row, etc.), the data is filled in sequence; the i -th element in the j -th row is:

[0079] (1)

[0080] where ;

[0081] For even rows i (such as the second row, the fourth row, etc.), the data is filled in reverse order; the i -th element in the j -th row is:

[0082] (2)

[0083] where ;

[0084] Further preferably, the wavelet transform (WT) is a mathematical tool for signal analysis and processing, which can decompose a signal into different frequency components; specifically, the two-dimensional discrete wavelet transform is adopted in this application, which decomposes the signal through a set of wavelet basis functions, so as to extract the local information of the signal; for a two-dimensional matrix, the two-dimensional discrete wavelet transform decomposes it into 4 components, including LL (low-frequency approximation part): containing the low-frequency information of the matrix, that is, the general contour of the matrix; LH (horizontal high-frequency part): containing the horizontal edge information of the matrix; HL (vertical high-frequency part): containing the vertical edge information of the matrix; HH (diagonal high-frequency part): containing the diagonal edge information of the matrix; the one-dimensional discrete wavelet transform (1D-DWT) convolves a one-dimensional signal through a wavelet function for a one-dimensional signal, that is:

[0085] (3)

[0086] wherein, is the scaled and translated form of the wavelet basis function, that is:

[0087] (4)

[0088] wherein, is the scale parameter (controlling the scaling of the wavelet basis function), is the translation parameter (controlling the position of the wavelet basis function), is the mother wavelet function;

[0089] The two-dimensional discrete wavelet transform is an extension of the one-dimensional transform; in the two-dimensional case, the one-dimensional discrete wavelet transform is first applied to each row of the matrix, and then 1D-DWT is applied to each column; since the one-dimensional data is converted into a two-dimensional matrix by using the dimension conversion function, denoted as ; wherein, x and y respectively represent the spatial coordinates in the horizontal and vertical directions; for performing wavelet transform, that is:

[0090] (5)

[0091] wherein, is the two-dimensional wavelet basis function, obtained by scaling and translating the one-dimensional wavelet transform in the horizontal and vertical directions, that is:

[0092] (6)

[0093] wherein, is the scale parameter, controlling the scaling degree of the wavelet basis function; and are the translation parameters respectively, which control the positions of the wavelet basis functions in the horizontal and vertical directions, is the mother wavelet function;

[0094] Further preferably, the DWTSC-UNet denoising network includes an input layer, a hidden layer, and an output layer; among them, the hidden layer is composed of a discrete wavelet transform layer, two spatial attention mechanism layers, three convolutional blocks, two channel attention mechanism layers, three upsampling convolutional blocks, and two connection layers;

[0095] Further preferably, the two-dimensional convolutional layers in the three convolutional blocks adopt convolutional kernels with different numbers of channels and the same size; 4-channel 3 3 convolutional path: Use 4 channels, with a size of 3 3 to extract the low-level features of the image; 32-channel 3 3 convolutional path: Use 32 channels, with a size of 3 3 to further abstract and extract higher-level features; 64-channel 3 3 convolutional path: Use 64 channels, with a size of 3 3 to extract more complex high-level features;

[0096] 2 2 max pooling path: Use 2 2 pooling window, stride 2 max pooling for downsampling to reduce the spatial dimensions (width and height);

[0097] Further preferably, in the three upsampling blocks, convolutional kernels with different numbers of channels and the same size are adopted, 128-channel 2 2 convolutional path: Use 128 channels, stride 2, size 2 2 to perform upsampling to restore the spatial resolution of the image;

[0098] 128-channel 2 2 convolutional path: Use 128 channels, stride 2, size 2 2 to perform upsampling to further restore the spatial resolution of the image;

[0099] 96-channel 2 2 convolutional path: Use 96 channels, stride 2, size 2 2 transposed convolution for upsampling to improve the spatial resolution of the image and restore the image size;

[0100] Further preferably, in the two spatial attention mechanism block layers, convolutional kernels of the same size are adopted, 7 7 convolutional path: Use 7 The 7×7 convolution captures spatial information;

[0101] Further preferably, in the two channel attention mechanism blocks, two 1×1 convolutions are used. The first 1×1 convolution reduces the number of input channels, and the second 1×1 convolution restores the number of channels; The first 1×1 convolution reduces the number of input channels, and the second 1×1 convolution restores the number of channels; The 1×1 convolution restores the number of channels;

[0102] The architecture of the denoising network is the improved UNet network, namely the DWT-CSUNet denoising network. The improvement compared with the traditional UNet network is that the DWT-CSUNet denoising network combines the advantages of discrete wavelet transform, spatial attention mechanism, channel attention mechanism and traditional U-shaped structure network. In this application, discrete wavelet transform is introduced. First of all, through multi-scale feature extraction, discrete wavelet transform can provide both high-frequency and low-frequency information of the signal, which helps the network extract richer hierarchical information from the data. Secondly, it can perform weighted learning on different frequency components of the image, improve the sensitivity to details and edges, and enhance the network's ability to express complex signals. Then, as a feature preprocessing method, discrete wavelet transform optimizes the data before network training, reduces the computational complexity of the network during the training process, and accelerates network training. Finally, in complex tasks, discrete wavelet transform helps the neural network capture a wider range of patterns by extracting multi-scale features of the signal, enhancing the generalization ability of the model. In addition, this application introduces a channel attention mechanism module and a spatial attention mechanism module. The internal structure of the module is simple, including a small number of convolution, pooling and feature fusion operations, with low computational complexity, saving computational resources. The channel attention mechanism is used to weight different channels (i.e., different dimensions of the feature map), assign a larger weight to important channels to highlight important features and enhance the network's attention to key features; assign a smaller weight to unimportant channels to suppress unimportant features, thereby reducing redundant information. The spatial attention mechanism weights different spatial positions of the feature map, which helps the network focus on important regions or targets in the image, ignoring the background or irrelevant parts. Secondly, by assigning higher weights to specific spatial regions, more detailed spatial information can be captured, further enhancing the denoising accuracy of the network. In controlled-source electromagnetic data denoising, this means that DWTSC-UNet can observe and understand the characteristics of controlled-source electromagnetic signals in more detail, which is very beneficial for accurately removing noise from controlled-source electromagnetic data because the noise may exist at different scales.

[0103] The UNet network plays a crucial role in the DWTSC-UNet network. The overall DWTSC-UNet network adopts an "encoder-decoder" model. The encoder (downsampling part) gradually reduces the image resolution through convolution and pooling operations, extracting increasingly abstract features. The decoder (upsampling part) gradually restores the spatial resolution and image size of the image. At the same time, the UNet network introduces skip connections, directly transmitting the feature maps in the encoder stage to the corresponding layers in the decoder stage. By directly transmitting high-resolution feature maps, it avoids the loss of spatial information caused by pooling operations, thus helping the model better process information such as edges, details, and textures in the image. The skip connections enable the gradient to be more easily propagated to the early layers of the network during backpropagation, thereby accelerating the training process and avoiding the problem of gradient vanishing. In addition, the advantage of the DWTSC-UNet network lies in its ability to make full use of the advantages of the UNet backbone network, discrete wavelet transform, and attention mechanism module. By combining the UNet backbone network, discrete wavelet transform, and attention mechanism module, the DWTSC-UNet network can simultaneously capture important information, extract local information, and multi-scale features, thus understanding controlled-source electromagnetic data more comprehensively. This method of comprehensively using different modules makes the DWTSC-Unet network more powerful and effective in extracting important features, retaining detail features, adapting to, and expressing complex signals. Therefore, the DWTSC-UNet network in this application constructs a structure that is more conducive to extracting the features of controlled-source electromagnetic data and suppressing the noise of controlled-source electromagnetic data in accordance with the structure of the UNet backbone network, combined with the attention mechanism module and discrete wavelet transform.

[0104] Based on the multi-scale feature extraction ability of discrete wavelet transform, the feature focusing characteristics of the attention mechanism, and the unique advantages of the UNet network in processing two-dimensional data, this application combines discrete wavelet transform, the attention mechanism, and the UNet network to optimize the UNet network structure, improve the performance of the network, and introduce the improved UNet network to solve the problem of strong noise suppression in controlled-source electromagnetic, thereby improving the noise suppression accuracy.

[0105] Further preferably, the discrete wavelet transform layer in the DWTSC-Unet denoising network uses a wavelet basis type of "db5" wavelet basis, that is, the Daubechies wavelet basis, and the boundary processing mode used is the 'periodic' boundary processing, that is, the periodic boundary condition.

[0106] Embodiment 1

[0107] The present application provides a method and system for suppressing strong interference in controlled-source electromagnetic data based on a SC-UNet (DWTSC-UNet) network optimized by dimension conversion and discrete wavelet transform. It introduces a deep learning method optimized by dimension conversion and discrete wavelet transform to solve the method for suppressing strong interference in controlled-source electromagnetic data, which specifically includes the following steps:

[0108] Step 1: Obtain the controlled-source electromagnetic data and segment it using a time window, and take each data segment as a sample.

[0109] Step 2: For typical actual noisy data segments, use the dictionary learning (SISC) algorithm for data denoising to obtain a reliable denoising effect, and obtain a reliable noise-free state of the noisy data segments to produce a high-quality one-dimensional sample library for the denoising model.

[0110] As Figure 2 shown, the embodiment of the present application provides a one-dimensional denoising sample library, which shows 4 typical samples. The one-dimensional sample library contains noise samples of different types and amplitudes; the noise samples include Gaussian noise signals, impulse noise signals, square wave noise signals, and mixed noise signals; in order to enable the present application to effectively distinguish useful signals and noise and ensure that the network has good generalization ability, it is necessary to provide as many useful signals and noise samples as possible during the training process so that the network can learn more features; the number of sample pairs contained in the sample library of the present application is 48,000 pairs, among which the number of sample pairs in the training set and the validation set are 42,000 pairs and 6,000 pairs respectively, and the ratio of the training set to the validation set is 7:1.

[0111] In the embodiment of the present application, before model training, the sample data is normalized to reduce the error introduced by the difference in sample amplitudes:

[0112]

[0113] where is the i th normalized sample; in other feasible embodiments, no specific limitation is made on this.

[0114] Step 3: Segment the one-dimensional sample library obtained in step 2, and use the dimension conversion function to perform dimension conversion on each sample respectively to convert the one-dimensional data into a two-dimensional matrix, that is, obtain a two-dimensional sample library.

[0115] As Figure 3 is a schematic diagram of dimension conversion by the dimension conversion function. The length of the sampling point sequence of each sample is 1600, and the sequence is converted into a matrix shape of 40×40 using the dimension conversion function.

[0116] In the embodiment of this application, the dimension conversion function performs dimension conversion in an "S" shape, that is, for odd rows (such as the first row, the third row, etc.), the data is filled in sequence; for even rows (such as the second row, the fourth row, etc.), the data is filled in reverse order;

[0117] Step 4: Construct a DWTSC-UNet denoising network, and input the constructed two-dimensional sample library into the DWTSC-UNet denoising network for training to obtain a controllable-source electromagnetic data denoising model;

[0118] As Figure 4 shown in the DWTSC-UNet denoising network, which consists of three parts: an input layer, a hidden layer, and an output layer;

[0119] In the embodiment of this application, the input layer uses a dimension conversion function to perform dimension transformation, converting the 1×1600 size to 1×40×40, and adding a batch size of 256 at the same time. The shape of the input layer is (256, 1, 40, 40);

[0120] In the embodiment of this application, the hidden layer consists of a discrete wavelet transform layer, a first shape reconstruction layer, a first convolutional block, a first spatial attention mechanism module, a second convolutional block, a second spatial attention mechanism module, a third convolutional block, a first upsampling block, a second shape reconstruction layer, a first skip connection layer, a first channel attention mechanism layer, a second upsampling block, a second skip connection layer, a second channel attention mechanism module, and a third upsampling block, a total of fifteen parts; the calculation processes of the discrete wavelet transform layer, convolutional block, spatial attention mechanism module, channel attention mechanism module, upsampling block, skip connection layer, and shape reconstruction layer in the embodiment of this application are as follows:

[0121] The steps of the discrete wavelet transform layer are as follows: First, the input shape is (256, 1, 40, 40). Using discrete wavelet transform, each image is decomposed into a low-frequency part (LL) and three high-frequency parts (LH, HL, HH), that is, decomposed into 4 channels. Then, through the "db5" wavelet basis and the 'periodic' boundary processing mode, the image shape is converted to (256, 4, 24, 24), and the result of the discrete wavelet transform layer is output;

[0122] The steps of the first shape reconstruction layer are as follows: First, input the result of the discrete wavelet transform layer, and use the bilinear interpolation method to adjust the output size of the discrete wavelet transform layer to the same size as the input layer, that is, (256, 4, 40, 40), and output the result of the shape reconstruction layer;

[0123] The steps of the first convolutional block are as follows: First, input the result of the first shape reconstruction layer, and use 32 filters with a size of 3 3. A convolutional kernel with a stride of 1 and a padding of 1 performs a two-dimensional convolution on the result of the first shape reconstruction layer and is activated using the "ReLU" activation function. Second, a batch normalization layer (BatchNormalization) is used to standardize the data for each channel. Third, a pooling window size of 2 2. A pooling kernel with a stride of 2 performs two-dimensional max pooling on the standardized result, reducing the spatial dimensions by half, i.e., (256, 32, 20, 20), and outputs the result of the first convolutional block.

[0124] The steps of the first spatial attention mechanism module are as follows: First, the result of the first convolutional block is input, and average pooling is used to calculate the channel average for each position in the result of the first convolutional block, with the result shape being (256, 1, 20, 20). Second, max pooling is used to calculate the channel maximum for each position in the result of the first convolutional block, with the result shape being (256, 1, 20, 20). Third, the results of average pooling and max pooling are concatenated along the channel dimension, with the result shape being (256, 2, 20, 20). Fourth, 2 convolutional kernels of size 7 7. A convolutional kernel with a stride of 1 and a padding of 3 performs a two-dimensional convolution on the concatenated result, with the result shape being (256, 1, 20, 20), and is activated using the "Sigmoid" activation function. Finally, the result of the first convolutional block is multiplied element-wise with the result after "Sigmoid" activation, with the result shape being (256, 32, 20, 20), and the result of the first spatial attention mechanism module is output.

[0125] The steps of the second convolutional block are as follows: First, the result of the first spatial attention mechanism module is input, and 64 convolutional kernels of size 3 3. A convolutional kernel with a stride of 1 and a padding of 1 performs a two-dimensional convolution on the result of the first spatial attention mechanism module and is activated using the "ReLU" activation function. Second, a batch normalization layer (BatchNormalization) is used to standardize the data for each channel. Third, a pooling window size of 2 2. A pooling kernel with a stride of 2 performs two-dimensional max pooling on the standardized result, reducing the spatial dimensions by half, i.e., (256, 64, 10, 10), and outputs the result of the second convolutional block.

[0126] The steps of the second spatial attention mechanism module are as follows: First, input the result of the second convolutional block, and use average pooling to calculate the channel average value of each position in the result of the second convolutional block, with the result shape being (256, 1, 10, 10); Second, use max pooling to calculate the channel maximum value of each position in the result of the second convolutional block, with the result shape being (256, 1, 10, 10); Third, concatenate the results of average pooling and max pooling along the channel dimension, with the result shape being (256, 2, 10, 10); Fourth, use two convolutional kernels with a size of 7 7, a stride of 1, and a padding of 3 to perform two-dimensional convolution on the concatenated result, with the result shape being (256, 1, 10, 10), and use the "Sigmoid" activation function for activation; Finally, multiply the result of the second convolutional block element-wise with the result after "Sigmoid" activation, with the result shape being (256, 64, 10, 10), and output the result of the second spatial attention mechanism module;

[0127] The steps of the third convolutional block are as follows: First, input the result of the second spatial attention mechanism module, and use 128 convolutional kernels with a size of 3 3, a stride of 1, and a padding of 1 to perform two-dimensional convolution on the result of the second spatial attention mechanism module, and use the "ReLU" activation function for activation; Second, use the batch normalization layer (BatchNormalization) to standardize the data of each channel; Third, use a pooling kernel with a window size of 2 2 and a stride of 2 to perform two-dimensional max pooling on the standardized result, reducing the spatial size by half, that is, (256, 128, 5, 5), and output the result of the third convolutional block;

[0128] The steps of the first upsampling block are as follows: First, input the result of the third convolutional block, and use 64 convolutional kernels with a size of 3 3 and a stride of 2 to perform two-dimensional transposed convolution on the result of the third convolutional block, expanding the spatial size by two times, that is, (256, 64, 10, 10), and output the result of the upsampling block;

[0129] The steps of the second shape reconstruction layer are as follows: First, input the result of the first upsampling block, and use the bilinear interpolation method to adjust the output size of the upsampling block to the same size as the input layer, that is, (256, 64, 10, 10), and output the result of the second shape reconstruction layer;

[0130] The steps of the first skip connection layer are as follows: First, input the results of the first shape reconstruction layer and the first upsampling block, concatenate the results of the second shape reconstruction layer and the first upsampling block, with the result shape being (256, 128, 10, 10), and output the result of the first skip connection layer;

[0131] The steps of the first channel attention mechanism layer are as follows: First, input the result of the first skip connection layer, and use average pooling to perform spatial average pooling on the result of the first skip connection layer, that is, pool the height and width dimensions, and the result shape is (256, 128, 1, 1); Second, use max pooling to perform max pooling on the result of the first skip connection layer in the height dimension, and the result shape is (256, 128, 1, 10), and then perform max pooling on the result of max pooling in the height dimension in the width dimension, and the result shape is (256, 128, 1, 1); Third, concatenate the average pooling and max pooling results in the channel dimension, and the result shape is (256, 256, 1, 1); Fourth, use 256 convolutional kernels with a size of 1 1. Perform two-dimensional convolution on the concatenated result using a convolutional kernel with a size of 1 and a stride of 1, and the result shape is (256, 8, 1, 1), and use the "ReLU" activation function for activation; Fifth, use 128 convolutional kernels with a size of 1 1. Perform two-dimensional convolution on the result after "ReLU" activation using a convolutional kernel with a size of 1 and a stride of 1, and the result shape is (256, 128, 1, 1), and use the "Sigmoid" activation function for activation; Finally, multiply the result of the first skip connection layer by the result after "Sigmoid" activation, and the result shape is (256, 128, 10, 10), and output the result of the first channel attention mechanism layer;

[0132] The steps of the second upsampling block are as follows: Input the result of the first channel attention mechanism layer, and use 64 convolutional kernels with a size of 3 3. Perform two-dimensional transposed convolution on the result of the first channel attention mechanism layer using a convolutional kernel with a size of 3 and a stride of 2 to double the spatial size, that is, (256, 64, 20, 20), and output the result of the second upsampling block;

[0133] The steps of the second skip connection layer are as follows: First, input the results of the first convolutional block and the second upsampling block, concatenate the results of the first convolutional block and the second upsampling block, and the result shape is (256, 96, 20, 20), and output the result of the second skip connection layer;

[0134] The steps of the second channel attention mechanism module are as follows: First, input the result of the second skip connection layer, and use average pooling to perform spatial average pooling on the result of the first skip layer, that is, pool the height and width dimensions, and the result shape is (256, 96, 1, 1); Second, use max pooling to perform max pooling on the result of the second skip connection layer in the height dimension, and the result shape is (256, 96, 1, 20), and then perform max pooling on the result of max pooling in the height dimension in the width dimension, and the result shape is (256, 96, 1, 1); Third, concatenate the average pooling and max pooling results in the channel dimension, and the result shape is (256, 192, 1, 1); Fourth, use 256 convolutional kernels with a size of 1 1. Perform two-dimensional convolution on the concatenated result using a convolutional kernel with a size of 1 and a stride of 1, and the result shape is (256, 6, 1, 1), and use the "ReLU" activation function for activation; Fifth, use 128 convolutional kernels with a size of 1 1. Perform two-dimensional convolution on the result after "ReLU" activation using a convolutional kernel with a size of 1 and a stride of 1, and the result shape is (256, 96, 1, 1), and use the "Sigmoid" activation function for activation; Finally, multiply the result of the first skip connection layer by the result after "Sigmoid" activation, and the result shape is (256, 96, 20, 20), and output the result of the second channel attention mechanism layer;

[0135] The steps of the third upsampling block are as follows: Input the result of the second channel attention mechanism layer, and use 64 with a size of 3 3. Perform two-dimensional transposed convolution on the result of the second channel attention mechanism layer using a convolutional kernel with a size of 3 and a stride of 2 to double the spatial size, that is, (256, 1, 40, 40), and output the result of the third upsampling block;

[0136] It should be noted that the number of the above upsampling blocks, the number of spatial attention mechanism modules, the number of channel attention mechanism modules, the number of skip connection layers, the number of convolutional kernels, and the size are set based on the training effect of the model. Therefore, the above embodiments are only for illustrative purposes. Without departing from the concept of the present application, the number of sampling blocks, the number of spatial attention mechanism modules, the number of channel attention mechanism modules, the number of skip connection layers, the size and number of convolutional kernels can be adjusted;

[0137] Further preferably, the Adam optimizer is selected during the training process of the denoising model. Among them, the training and validation batch sizes are both 256, the initial learning rate is 1×10 -5 ⁻⁴, and the learning rate decay method is used, and the learning rate decays by 0.1 every 10 training epochs, and a total of 500 times of training are performed; In other feasible embodiments, no specific limitation is made on this, and other optimizers can be selected;

[0138] As Figure 5 is a schematic diagram of multi-scale feature decomposition by discrete wavelet transform. The two-dimensional matrix is decomposed into four frequency components through discrete wavelet transform, including a low-frequency part (LL) and three high-frequency parts (LH, HL, HH);

[0139] Figure 6 is the loss change during the training process of the DWTSC-UNet denoising network model. Among them, the solid line represents the loss change of the validation set during the model training process, and the dotted line represents the loss change of the training set after the model is trained with the training set; Judging from the curve form, as the number of training times increases, the loss value of the model gradually decreases and finally stabilizes, indicating that the data features learned by the model tend to grow to convergence, the adaptability of the model to the data is increasing, and the error of the model is decreasing;

[0140] Step 5: Use each data segment in Step 1 as a sample and label the class label of the sample to construct a classification sample library; The class label indicates that the sample is a noise data segment or a high-quality data segment;

[0141] As Figure 7 is the classification sample library provided by this application, showing 16 typical samples in the sample library; The first column is the measured relatively high-quality sample; The second column is the measured noisy sample; The third column is the simulated noise-free sample; The fourth column is the simulated noisy sample; It can be seen that the noise of the controlled-source signal noise has many types such as Gaussian noise, harmonic noise, square-wave noise, impulse-like noise, and baseline drift noise;

[0142] Step 6: Construct an IncepTCN classification network, and input the classification sample library into the IncepTCN classification network for training to obtain a controlled-source electromagnetic data classification model;

[0143] Figure 8 is a schematic diagram of the structure of the IncepTCN data classification model constructed by this application. The IncepTCN data classification model consists of an input layer, a hidden layer, and an output layer; The shape of the input layer is: the number of channels multiplied by the sample length, that is, 1 1600;

[0144] In the embodiment of this application, the hidden layer consists of a convolutional layer, a first multi-scale convolutional module, a first time-domain convolutional block, a first max-pooling layer, a second multi-scale convolutional module, a second time-domain convolutional block, a second max-pooling layer, a third multi-scale convolutional module, a third time-domain convolutional block, a third max-pooling layer, a fourth multi-scale convolutional module, a fourth time-domain convolutional block, and a fourth max-pooling layer, a total of thirteen parts; The calculation processes of the convolutional layer, multi-scale convolutional module, time-domain convolutional block, and max-pooling layer in the embodiment of this application are:

[0145] The steps of the convolutional layer are as follows: Use 16 convolutional kernels of size 16 ×1 to perform one-dimensional convolution on the training samples, and use "ReLU" activation to output the result of the convolutional layer;

[0146] The steps of the first multi-scale convolutional module are as follows: First, input the result of the convolutional layer. The four paths of the first multi-scale convolutional module are as follows: 1 ×3 convolutional path: Use 1 ×3 convolution to extract local spatio-temporal features; 1 ×5 convolutional path: Use 1 ×5 convolution to extract medium-range spatio-temporal features; 1 ×7 convolutional path: Use 1 ×7 convolution to extract wider-range spatio-temporal features; 1 ×3 max-pooling path: Use 1 ×3 max-pooling to fuse features in the local range; Finally, connect the outputs of the four paths as the final output of the first multi-scale convolutional module to achieve multi-scale feature extraction and use it as the result of the first multi-scale convolutional module;

[0147] The steps of the first time-domain convolutional block are as follows: First, use 64 convolutional kernels of size 3 ×1 to perform one-dimensional causal dilated convolution and one-dimensional convolution on the result of the first multi-scale convolutional module respectively. The causal dilated convolution uses "ReLU" activation, and the one-dimensional convolution is not activated; Second, use 64 convolutional kernels of size 3 ×1 again to perform one-dimensional causal dilated convolution on the activation value of the first causal dilated convolution of the residual block; Third, sum the one-dimensional convolution of the residual block and the second non-activated one-dimensional causal dilated convolution of the residual block, and use "ReLU" activation to output the result of the first time-domain convolutional block;

[0148] The steps of the first max-pooling layer are as follows: First, input the result of the first time-domain convolutional block, and use max-pooling operation with a size of 2 and a stride of 2 for screening to output the result of the first max-pooling layer;

[0149] The steps of the second multi-scale convolutional module are as follows: First, input the result of the first max-pooling layer. The four paths of the second multi-scale convolutional module are as follows: 1 ×3 convolutional path: Use 1 ×3 convolution to extract local spatio-temporal features; 1 ×5 convolutional path: Use 1 ×5 convolution to extract medium-range spatio-temporal features; 1 ×7 convolutional path: Use 1 ×7 convolution to extract wider-range spatio-temporal features; 1 ×3 max-pooling path: Use 1 3 max pooling fuses the features in the local range; finally, the outputs of the four paths are concatenated as the final output of the second multi-scale convolution module, achieving multi-scale feature extraction and serving as the result of the second multi-scale convolution module;

[0150] The steps of the second temporal convolution block are as follows: First, use 64 convolutional kernels with a size of 3 × 1 to perform one-dimensional causal dilated convolution and one-dimensional convolution on the result of the second multi-scale convolution module respectively. The causal dilated convolution uses "ReLU" activation, and the one-dimensional convolution is not activated; Second, use 64 convolutional kernels with a size of 3 × 1 to perform one-dimensional causal dilated convolution on the activation value of the first causal dilated convolution of the residual block; Third, sum the one-dimensional convolution of the residual block and the second unactivated one-dimensional causal dilated convolution of the residual block, and use "ReLU" activation to output the result of the second temporal convolution block;

[0151] The steps of the second max pooling layer are as follows: First, input the result of the second temporal convolution block, and use max pooling operation with a size of 2 and a stride of 2 for screening to output the result of the max pooling layer;

[0152] The steps of the third multi-scale convolution module are as follows: First, input the result of the second max pooling layer. The four paths of the third multi-scale convolution module are as follows: 1 × 3 convolution path: Use 1 × 3 convolution to extract local spatio-temporal features; 1 × 5 convolution path: Use 1 × 5 convolution to extract medium-range spatio-temporal features; 1 × 7 convolution path: Use 1 × 7 convolution to extract wider-range spatio-temporal features; 1 × 3 max pooling path: Use 1 × 3 max pooling to fuse the features in the local range; finally, the outputs of the four paths are concatenated as the final output of the third multi-scale convolution module, achieving multi-scale feature extraction and serving as the result of the third multi-scale convolution module;

[0153] The steps of the third temporal convolution block are as follows: First, use 64 convolutional kernels with a size of 3 × 1 to perform one-dimensional causal dilated convolution and one-dimensional convolution on the result of the third multi-scale convolution module respectively. The causal dilated convolution uses "ReLU" activation, and the one-dimensional convolution is not activated; Second, use 64 convolutional kernels with a size of 3 × 1 to perform one-dimensional causal dilated convolution on the activation value of the first causal dilated convolution of the residual block; Third, sum the one-dimensional convolution of the residual block and the second unactivated one-dimensional causal dilated convolution of the residual block, and use "ReLU" activation to output the result of the third temporal convolution block;

[0154] The steps of the third max pooling layer are as follows: First, input the result of the third temporal convolutional block, and perform screening using a max pooling operation with a size of 2 and a stride of 2 to output the result of the third max pooling layer.

[0155] The steps of the fourth multi-scale convolutional module are as follows: First, input the result of the third max pooling layer. The four paths of the fourth multi-scale convolutional module are as follows: 1 3-convolution path: Use 1 3-convolutions to extract local spatio-temporal features; 1 5-convolution path: Use 1 5-convolutions to extract medium-range spatio-temporal features; 1 7-convolution path: Use 1 7-convolutions to extract wider-range spatio-temporal features; 1 3-max pooling path: Use 1 3-max pooling to fuse features in the local range; finally, connect the outputs of the four paths as the final output of the fourth multi-scale convolutional module to achieve multi-scale feature extraction and use it as the result of the fourth multi-scale convolutional module;

[0156] The steps of the fourth temporal convolutional block are as follows: First, use 64 convolutional kernels with a size of 3 ×1 to perform one-dimensional causal dilated convolution and one-dimensional convolution on the result of the third multi-scale convolutional module respectively. The causal dilated convolution uses the "ReLU" activation, and the one-dimensional convolution is not activated; Second, use 64 convolutional kernels with a size of 3 ×1 again to perform one-dimensional causal dilated convolution on the activation value of the first causal dilated convolution of the residual block; Third, sum the one-dimensional convolution of the residual block and the second unactivated one-dimensional causal dilated convolution of the residual block, and use "ReLU" activation to output the result of the fourth temporal convolutional block;

[0157] The steps of the fourth max pooling layer are as follows: First, input the result of residual module 12, and perform screening using a max pooling operation with a size of 2 and a stride of 2 to output the result of max pooling layer 13.

[0158] It should be noted that the number of the above multi-scale convolutional modules, the number of residual modules, the number of convolutional kernels, and the size are set based on the training effect of the model. Therefore, the above embodiments are only for illustrative purposes. Without departing from the concept of the present application, the number of multi-scale convolutional modules, the number of residual modules, the number of convolutional kernels, and the size can be adjusted;

[0159] Finally, the output layer of IncepTCN consists of 1 Flatten layer and 1 fully connected layer; the Flatten layer flattens the result of the max pooling layer into 1 One-dimensional sequence of 96,000; the fully connected layer contains 2 neurons and is activated using the "Softmax" function; the output layer outputs the probability of the category, thereby obtaining the classification result of the sample, that is, noise or high quality;

[0160] During the training process of the classification model in the embodiments of the present application, the Adam optimizer is selected. Among them, the training and validation Batchsize are both 256, and the initial learning rate is 1×10 -5 , and a total of 30 epochs are trained; in other feasible embodiments, no specific limitation is made on this, and other optimizers can be selected;

[0161] Figure 9 The figure shows the accuracy and loss changes during the training process of the IncepTCN classification network model. Among them, the solid line represents the changes in accuracy and loss of the validation set during the model training process, and the dotted line represents the changes in accuracy and loss of the training set after the model is trained on the training set; from the curve form, as the number of training times increases, the accuracy of the model is gradually increasing, and the loss value of the model is also gradually decreasing, and finally it also tends to be stable, which indicates that the data features learned by the model show a trend of growth to convergence, the adaptability of the model to the data is increasing, and the error of the model is decreasing;

[0162] Step 7: Input the controllable-source electromagnetic data to be processed, segment it first, and then input it into the controllable-source electromagnetic data classification model to obtain the classification result of each data segment, and retain the data segments identified as high quality by the model;

[0163] Step 8: Input the controllable-source electromagnetic data segments identified as noisy by the classification model in Step 7 into the controllable-source electromagnetic data denoising model to obtain the data denoising results of each data segment;

[0164] Step 9: Use the dimension conversion function in Step 3 to perform an inverse dimension transformation on the denoising result obtained in Step 8, and convert the two-dimensional data into a one-dimensional time series;

[0165] Step 10: Combine the high-quality data retained in Step 7 and the data after the dimension conversion of the denoised data in Step 9 to obtain a complete controllable-source electromagnetic denoised sequence.

[0166] Figure 10It is the confusion matrix of the IncepTCN classification model. Here, the abscissa is the result predicted by the IncepTCN classification model, and the ordinate is the manually marked result; the label equal to 0 indicates high quality, and the label equal to 1 indicates noise. The upper left, upper right, lower left, and lower right parts of each subgraph represent the numbers of PT samples, NF samples, PF samples, and NT samples respectively; the darker the color, the larger the value; from left to right are the cases of signal-dominated, signal-noise balanced, and noise-dominated; in the case where high-quality samples dominate, the IncepTCN classification model achieved excellent results, with an F1 score exceeding 0.99; when the numbers of high-quality samples and noise samples are balanced, the accuracy of the IncepTCN classification model exceeds 0.99; when noise samples dominate, the IncepTCN classification model also shows good performance, with an F1 score exceeding 0.95;

[0167] Figure 11 It is the classification effect diagram of the IncepTCN classification model. Here, from top to bottom are the cases of signal-dominated, signal-noise balanced, and noise-dominated, and it can be seen from Figure 11 that the recognition effect of the IncepTCN classification model is accurate, and the noisy segments are all recognized.

[0168] Figure 12 It is the denoising effect diagram of the DWTSC-UNet model; here, the dotted line represents the synthesized noise samples, and the solid line represents the result after denoising by the DWTSC-UNet model. It can be seen from Figure 12 that the noise is effectively removed by the DWTSC-UNet model, and the overall curve becomes smooth, and no obvious noise can be seen;

[0169] Figure 13 They are the apparent resistivity curves and U / I curves before and after corresponding data processing; here, the dotted line represents the apparent resistivity and U / I curves calculated from the time series of the synthesized noise signal, and the solid line represents the apparent resistivity and U / I curves calculated from the time series after being processed by the method of the present application. From left to right are the noise-containing signals with signal-to-noise ratios of 5 dB, -5 dB, -10 dB, and -15 dB; it can be seen from Figure 13 that when the signal-to-noise ratio is 5 dB, the original noise curve presents a relatively smooth curve with a slight distortion at 3 Hz. When the signal-to-noise ratios are -55 dB, -10 dB, and -15 dB, the original noise curve shows significant distortion in the frequency range below 10 Hz. After being processed by the method of the present application, the distorted curves become smooth and continuous.

[0170] Embodiment 2

[0171] The present application provides a controllable-source electromagnetic data strong interference suppression system, including:

[0172] A data classification module, which is used to input the controllable-source electromagnetic data to be processed into a controllable-source electromagnetic data classification model to obtain the first noiseless controllable-source electromagnetic data and the noisy controllable-source electromagnetic data;

[0173] A data denoising module, which is used to convert the noisy controllable-source electromagnetic data into two-dimensional image data through a dimension conversion function and input it into a DWTSC-UNet denoising network model to obtain a denoised data result;

[0174] A dimension conversion module, which is used to perform an inverse dimension conversion on the denoised data result by using a dimension conversion function to obtain the second noiseless controllable-source electromagnetic data;

[0175] A data combination module, which is used to combine the first noiseless controllable-source electromagnetic data and the second noiseless controllable-source electromagnetic data to obtain complete noiseless controllable-source electromagnetic data;

[0176] Among them, the DWTSC-UNet denoising network model is a U-shaped structure network that combines discrete wavelet transform, spatial attention mechanism, and channel attention mechanism;

[0177] The discrete wavelet transform is used to provide high-frequency signals and low-frequency signals, perform weighted learning on different frequency components of the two-dimensional image data, and extract multi-scale features from the two-dimensional image data; the channel attention mechanism is used to weight different dimensions of the feature map; the spatial attention mechanism is used to weight different spatial positions of the feature map.

[0178] Further preferably, the controllable-source electromagnetic data strong interference suppression system further includes a denoising model training module, which includes:

[0179] An actual measured noise signal acquisition unit, which is used to select a sample of noisy controllable-source electromagnetic data from the actual measured data, perform denoising by using a dictionary learning method, and separate the actual measured noise signal;

[0180] A one-dimensional denoising sample library construction unit, which is used to add the actual measured noise signal and the simulated noise signal to the noiseless controllable-source electromagnetic data to form noisy controllable-source electromagnetic data, form a sample pair with the noiseless controllable-source electromagnetic data sample, and construct a one-dimensional denoising sample library;

[0181] A dimension conversion unit, which segments the one-dimensional denoising sample library and uses a dimension conversion function to perform dimension conversion on each sample to obtain a two-dimensional image sample library;

[0182] A model training unit, which is used to input the two-dimensional image sample library into the DWTSC-UNet denoising network model for training.

[0183] Further preferably, the DWTSC-UNet denoising network model in the data denoising module includes an input layer, a hidden layer, and an output layer; the hidden layer includes: a discrete wavelet transform layer, a first shape reconstruction layer, a first convolutional block, a first spatial attention mechanism module, a second convolutional block, a second spatial attention mechanism module, a third convolutional block, a first upsampling block, a second shape reconstruction layer, a first skip connection layer, a first channel attention mechanism layer, a second upsampling block, a second skip connection layer, a second channel attention mechanism module, and a third upsampling block;

[0184] Among them, the discrete wavelet transform layer is used to decompose the two-dimensional image data into a low-frequency approximation part, a horizontal high-frequency part, a vertical high-frequency part, and a diagonal high-frequency part; the low-frequency approximation part contains the basic contour of the two-dimensional image data; the horizontal high-frequency part contains the horizontal edge information of the two-dimensional image data; the vertical high-frequency part contains the vertical edge information of the two-dimensional image data; the diagonal high-frequency part contains the diagonal edge information of the two-dimensional image data.

[0185] Further preferably, the discrete wavelet transform layer uses a wavelet basis type of db5 wavelet basis and a boundary processing mode of periodic boundary conditions.

[0186] Further preferably, the controlled-source electromagnetic data classification model in the data classification module is an IncepTCN classification network model.

[0187] In summary, compared with the prior art, the present application has the following advantages:

[0188] The present application introduces a dimension conversion method to convert the problem of controlled-source electromagnetic signal denoising into an image denoising problem. Using the proposed signal-to-image transformation method, the one-dimensional controlled-source electromagnetic time series is converted into two-dimensional data to retain the structural characteristics of the original controlled-source electromagnetic signal.

[0189] The present application introduces a method that combines discrete wavelet transform and UNet network, making full use of the advantages of discrete wavelet transform and UNet architecture. First, the discrete wavelet transform performs multi-scale feature extraction on the controlled-source electromagnetic data, extracting different frequency information, which helps the network better understand the content of the controlled-source electromagnetic data; then, combining the discrete wavelet transform and UNet architecture can perform more effective feature fusion between the encoder and decoder of the model, improving the performance of the network; in addition, UNet itself has good local feature capture ability, and the features extracted by the encoder part are directly passed to the decoder part using skip connections. The discrete wavelet transform can further strengthen this ability, especially at the detail level of the image, helping the model better capture the local structure information of the image; furthermore, the discrete wavelet transform can compress the high-dimensional representation of the image into a low-dimensional multi-scale representation, reducing the computational complexity and improving the processing efficiency of the network.

[0190] This application makes full use of the ability of the attention mechanism to improve the model's feature extraction ability, the multi-scale feature extraction ability of discrete wavelet transform, as well as the unique encoding-decoding structure, skip connections and effective integration ability of local and global information of the UNet network, and proposes a new denoising network, namely the DWTSC-UNet denoising network; without inspecting the effective signal, it realizes strong noise suppression of controlled-source electromagnetic data, and the calculated apparent resistivity curve of the denoised data is significantly improved, and the curve becomes smoother and more continuous.

[0191] After the model training is completed, the processes of data processing such as recognition and denoising are all fully automated by the computer, without any manual intervention, and there is no experience requirement for data processing operators. It not only eliminates the problem of subjective deviation caused by manually setting thresholds in traditional methods, but also improves the adaptability to different types of noise.

[0192] It can be understood that the detailed function implementation of each of the above units / modules can be referred to the introduction in the foregoing method embodiments, and will not be elaborated here.

[0193] It should be understood that the above system is used to execute the method in the above embodiments. The corresponding program modules in the system have similar implementation principles and technical effects as those described in the above method. The working process of the system can refer to the corresponding process in the above method, and will not be elaborated here.

[0194] Based on the method in the above embodiments, an embodiment of this application provides an electronic device, which may include: a processor (Processor), a communication interface (Communications Interface), a memory (Memory) and a communication bus. Among them, the processor, the communication interface, and the memory complete communication with each other through the communication bus. The processor can call the logical instructions in the memory to execute the method in the above embodiments.

[0195] Those skilled in the art can easily understand that the above is only a preferred embodiment of this application, and is not intended to limit this application. Any modifications, equivalent replacements, and improvements made within the spirit and principles of this application shall be included within the protection scope of this application.

Claims

1. A method for suppressing strong interference in controlled-source electromagnetic data, characterized in that It includes the following steps: Input the controllable source electromagnetic data to be processed into the controllable source electromagnetic data classification model to obtain the first noise-free controllable source electromagnetic data and the noise-containing controllable source electromagnetic data; Convert the noise-containing controllable source electromagnetic data into two-dimensional image data through a dimension conversion function and input it into the DWTSC-UNet denoising network model to obtain the denoising data result; Perform inverse dimension conversion on the denoising data result using the dimension conversion function to obtain the second noise-free controllable source electromagnetic data; Combine the first noise-free controllable source electromagnetic data and the second noise-free controllable source electromagnetic data to obtain the complete noise-free controllable source electromagnetic data; Among them, the DWTSC-UNet denoising network model is a U-shaped structure network that combines discrete wavelet transform, spatial attention mechanism and channel attention mechanism; The discrete wavelet transform is used to provide high-frequency signals and low-frequency signals, perform weighted learning on different frequency components of the two-dimensional image data, and perform multi-scale feature extraction from the two-dimensional image data; the channel attention mechanism is used to weight different dimensions of the feature map; the spatial attention mechanism is used to weight different spatial positions of the feature map.

2. The controllable source electromagnetic data strong interference suppression method according to claim 1, wherein The training method of the DWTSC-UNet denoising network model includes the following steps: Select the noise-containing controllable source electromagnetic data samples from the measured data, perform denoising using the dictionary learning method, and separate the measured noise signal; Add the measured noise signal and the simulated noise signal to the noise-free controllable source electromagnetic data to form the noise-containing controllable source electromagnetic data, form sample pairs with the noise-free controllable source electromagnetic data samples, and construct a one-dimensional denoising sample library; Segment the one-dimensional denoising sample library, perform dimension conversion on each sample using the dimension conversion function to obtain a two-dimensional image sample library; Input the two-dimensional image sample library into the DWTSC-UNet denoising network model for training.

3. The controllable source electromagnetic data strong interference suppression method according to claim 1 or 2, characterized in that The DWTSC-UNet denoising network model includes an input layer, a hidden layer and an output layer; The hidden layer includes: a discrete wavelet transform layer, a first shape reconstruction layer, a first convolutional block, a first spatial attention mechanism module, a second convolutional block, a second spatial attention mechanism module, a third convolutional block, a first upsampling block, a second shape reconstruction layer, a first skip connection layer, a first channel attention mechanism layer, a second upsampling block, a second skip connection layer, a second channel attention mechanism module and a third upsampling block.

4. The controllable source electromagnetic data strong interference suppression method according to claim 3, wherein The discrete wavelet transform layer is used to decompose the two-dimensional image data into a low-frequency approximation part, a horizontal high-frequency part, a vertical high-frequency part and a diagonal high-frequency part; the low-frequency approximation part contains the basic contour of the two-dimensional image data; the horizontal high-frequency part contains the horizontal edge information of the two-dimensional image data; the vertical high-frequency part contains the vertical edge information of the two-dimensional image data; the diagonal high-frequency part contains the diagonal edge information of the two-dimensional image data.

5. The controllable-source electromagnetic data strong interference suppression method according to claim 4, wherein The discrete wavelet transform layer uses the db5 wavelet basis type and the boundary processing mode is the periodic boundary condition.

6. The controllable-source electromagnetic data strong interference suppression method according to claim 1, wherein The controllable source electromagnetic data classification model is the IncepTCN classification network model.

7. A controllable source electromagnetic data strong interference suppression system, characterized in that, It includes: A data classification module, which is used to input the controllable source electromagnetic data to be processed into a controllable source electromagnetic data classification model to obtain first noise-free controllable source electromagnetic data and noise-containing controllable source electromagnetic data; A data denoising module, which is used to convert the noise-containing controllable source electromagnetic data into two-dimensional image data through a dimension conversion function and input it into a DWTSC-UNet denoising network model to obtain a denoising data result; A dimension conversion module, which is used to perform an inverse dimension conversion on the denoising data result by using a dimension conversion function to obtain second noise-free controllable source electromagnetic data; A data combination module, which is used to combine the first noise-free controllable source electromagnetic data and the second noise-free controllable source electromagnetic data to obtain complete noise-free controllable source electromagnetic data; Among them, the DWTSC-UNet denoising network model is a U-shaped structure network that combines discrete wavelet transform, spatial attention mechanism, and channel attention mechanism; The discrete wavelet transform is used to provide high-frequency signals and low-frequency signals, perform weighted learning on different frequency components of the two-dimensional image data, and extract multi-scale features from the two-dimensional image data; the channel attention mechanism is used to weight different dimensions of the feature map; the spatial attention mechanism is used to weight different spatial positions of the feature map.

8. The controllable-source electromagnetic data strong interference suppression system according to claim 7, characterized in that It also includes a denoising model training module, which includes: An actual measured noise signal acquisition unit, which is used to select a noise-containing controllable source electromagnetic data sample from the actual measured data, perform denoising by using a dictionary learning method, and separate the actual measured noise signal; A one-dimensional denoising sample library construction unit, which is used to add the actual measured noise signal and the simulated noise signal to the noise-free controllable source electromagnetic data to form noise-containing controllable source electromagnetic data, form a sample pair with the noise-free controllable source electromagnetic data sample, and construct a one-dimensional denoising sample library; A dimension conversion unit, which segments the one-dimensional denoising sample library and uses a dimension conversion function to perform dimension conversion on each sample to obtain a two-dimensional image sample library; A model training unit, which is used to input the two-dimensional image sample library into the DWTSC-UNet denoising network model for training.

9. The controllable-source electromagnetic data strong interference suppression system according to claim 7 or 8, characterized in that, The DWTSC-UNet denoising network model in the data denoising module includes an input layer, a hidden layer, and an output layer; the hidden layer includes: a discrete wavelet transform layer, a first shape reconstruction layer, a first convolutional block, a first spatial attention mechanism module, a second convolutional block, a second spatial attention mechanism module, a third convolutional block, a first upsampling block, a second shape reconstruction layer, a first skip connection layer, a first channel attention mechanism layer, a second upsampling block, a second skip connection layer, a second channel attention mechanism module, and a third upsampling block; Among them, the discrete wavelet transform layer is used to decompose the two-dimensional image data into a low-frequency approximation part, a horizontal high-frequency part, a vertical high-frequency part, and a diagonal high-frequency part; the low-frequency approximation part contains the basic contour of the two-dimensional image data; the horizontal high-frequency part contains the horizontal edge information of the two-dimensional image data; the vertical high-frequency part contains the vertical edge information of the two-dimensional image data; the diagonal high-frequency part contains the diagonal edge information of the two-dimensional image data.

10. The controllable source electromagnetic data strong interference suppression system according to claim 9, wherein The discrete wavelet transform layer uses a wavelet basis type of db5 wavelet basis and a boundary processing mode of periodic boundary conditions.

Images